Team 12
PRAVAH
See upstream. Protect downstream.
A flood intelligence prototype for Nepal that combines satellite water mapping, exposure screening and a map dashboard, with U-Net training distributed across two GB10 nodes.
Dhrut Patel, Pitambar Pandey, Sachin Shrestha, Preshak Bhattarai, Bishal Adhikari
University of Wisconsin-Green Bay
1.81×
training speedup on two nodes
486.06 s down to 268.82 s for 500 steps.
90.4%
parallel efficiency
Two-GPU distributed data parallel against one GPU.
14.88
samples per second
Up from 8.23 on one node.
Figures are the team's own measurements from the hackathon weekend, as reported in their submission.
The problem
Floods and river blockages caused by landslides in the steep Himalayan watersheds of Nepal can threaten downstream roads, bridges, settlements and emergency access. The Rasuwa flood of August 26, 2026 on the Bhote Koshi and Trishuli rivers motivated the focus on upstream awareness.
What they built
An evidence-aware decision-support prototype in five layers: data from Sentinel imagery, terrain, historical hydrology and OpenStreetMap; preprocessing; a compact U-Net that classifies water against non-water pixels; geospatial analysis that screens roads and buildings for potential exposure; and delivery through FastAPI endpoints and a Streamlit map dashboard.
The architecture supports a historical replay that separates evidence available before a chosen cutoff from what was observed afterward.
How they used the cluster
Fuzzball workflows schedule model training on the GB10 GPUs. The benchmark trained a new baseline U-Net twice with the same data manifest, seed, learning rate, 500 optimizer steps and global batch size of eight: once on one GPU, and once with distributed data parallel on two.
What the numbers do and do not show
The matched benchmark shows a 44.7% reduction in measured training time. The team is specific that this is training performance for the tested workload, not a speedup for satellite inference, API latency or the full response. Validation F1 was 0.8192 on one GPU and 0.8434 on two, which the team treats as checkpoint variation, not evidence that distributed training improves accuracy.
Segmentation and map overlays alone do not establish advance prediction, flood arrival time or confirmed damage. Next steps are to verify outputs against independent references for Nepal and enforce strict timestamp cutoffs.